Fetching the paper…

Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits · Around